An Efficient and Precise Training Data Construction Framework for Process-supervised Reward Model in Mathematical Reasoning

TL;DR

EpicPRM framework constructs high-quality math reasoning dataset Epic50k using adaptive binary search, enhancing PRM performance.

cs.CL 🔴 Advanced 2025-03-04 5 views
Wei Sun Qianlong Du Fuwei Cui Jiajun Zhang
mathematical reasoning large language models reward model dataset construction adaptive algorithm

Key Findings

Methodology

The EpicPRM framework quantifies the contribution of intermediate reasoning steps and uses an adaptive binary search algorithm to improve annotation precision and efficiency. It combines multiple large language models to generate diverse reasoning chains and annotates intermediate steps by quantifying their contribution, identifying the first erroneous step in each chain.

Key Results

  • Epic50k dataset contains 50k annotated intermediate steps, with PRM trained on Epic50k showing significantly superior performance compared to other public datasets, reducing annotation costs by 64.39%.
  • Epic50k outperforms Math-Shepherd and PRM800k across multiple models and sampling temperatures, especially in more complex problems.
  • Epic50k demonstrates strong performance in PROCESSBENCH evaluation, particularly excelling in Olympiad-level problems.

Significance

This research significantly enhances the efficiency and precision of mathematical reasoning, particularly in training large language models. By reducing dataset size without sacrificing performance, it offers a more economical way to construct training data, addressing the high cost and low quality issues of existing manual and automatic annotation methods.

Technical Contribution

The EpicPRM framework significantly improves annotation efficiency and data quality through an adaptive binary search algorithm and contribution quantification method. Compared to existing methods, EpicPRM maintains or even surpasses performance while reducing dataset size, providing new theoretical guarantees and engineering possibilities.

Novelty

EpicPRM is the first to apply an adaptive binary search algorithm to construct mathematical reasoning datasets, improving annotation precision by quantifying step contributions. This approach maintains high performance while reducing dataset size, distinguishing itself from existing manual and automatic annotation methods.

Limitations

  • EpicPRM may still require significant computational resources to ensure annotation precision when handling very complex problems.
  • The framework's applicability and scalability in specific domains need further validation.

Future Work

Future work includes extending the EpicPRM framework to support reasoning tasks in more domains and optimizing the use of computational resources. Additionally, research will explore further improvements in annotation efficiency and dataset quality.

AI Executive Summary

In the field of mathematical reasoning, existing large language models face high costs and low quality in dataset construction. The EpicPRM framework significantly improves annotation efficiency and precision through an adaptive binary search algorithm and contribution quantification method. This framework generates the Epic50k dataset, containing 50k annotated intermediate steps, significantly enhancing PRM performance.

EpicPRM combines multiple large language models to generate diverse reasoning chains and annotates intermediate steps by quantifying their contribution, identifying the first erroneous step in each chain. This approach reduces dataset size while maintaining high performance, offering a more economical way to construct training data.

Experimental results show that Epic50k outperforms Math-Shepherd and PRM800k across multiple models and sampling temperatures, especially in more complex problems. Epic50k demonstrates strong performance in PROCESSBENCH evaluation, particularly excelling in Olympiad-level problems. Future work includes extending the EpicPRM framework to support reasoning tasks in more domains and optimizing the use of computational resources.

Deep Analysis

Background

In mathematical reasoning with large language models, the quality and scale of datasets significantly impact model performance. Traditional manual annotation methods, while high quality, are costly and hard to scale. Automatic annotation methods, while reducing costs, often lack quality. The EpicPRM framework offers an efficient and precise way to construct datasets using an adaptive binary search algorithm and contribution quantification method.

Core Problem

Existing methods for constructing mathematical reasoning datasets face high costs and low quality. Manual annotation methods, while high quality, are difficult to scale to large datasets. Automatic annotation methods, while reducing costs, often lack quality, necessitating larger datasets to compensate for the lack of quality.

Innovation

The EpicPRM framework improves annotation efficiency and data quality through an adaptive binary search algorithm and contribution quantification method. The adaptive binary search algorithm dynamically adjusts the starting position and sampling count based on problem difficulty, significantly reducing computational costs. The contribution quantification method improves annotation precision by quantifying the contribution of intermediate steps.

Methodology

  • �� Use multiple large language models to generate diverse reasoning chains.
  • �� Annotate intermediate steps by quantifying their contribution.
  • �� Use an adaptive binary search algorithm to identify the first erroneous step in each chain.
  • �� Dynamically adjust the starting position and sampling count to improve efficiency.

Experiments

Experiments use the Epic50k dataset, containing 50k annotated intermediate steps. Multiple large language models are evaluated at different sampling temperatures, comparing the performance of Epic50k with Math-Shepherd and PRM800k. Results show that Epic50k outperforms other datasets across multiple models and sampling temperatures.

Results

Epic50k outperforms Math-Shepherd and PRM800k across multiple models and sampling temperatures, especially in more complex problems. Epic50k demonstrates strong performance in PROCESSBENCH evaluation, particularly excelling in Olympiad-level problems.

Applications

The EpicPRM framework can be used to construct high-quality mathematical reasoning datasets, suitable for training large language models. This framework offers a more economical way to construct training data, suitable for both academia and industry.

Limitations & Outlook

EpicPRM may still require significant computational resources to ensure annotation precision when handling very complex problems. The framework's applicability and scalability in specific domains need further validation. Future work includes extending the EpicPRM framework to support reasoning tasks in more domains and optimizing the use of computational resources.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen. EpicPRM is like a smart chef assistant that tells you what to do at each step and corrects you when you make a mistake. This assistant decides whether a correction is needed based on the importance of each step, just like some steps in a complex recipe are more crucial than others. This way, EpicPRM helps you complete the dish faster and more accurately without wasting too many ingredients and time.

ELI14 Explained like you're 14

Imagine you're playing a puzzle game, and EpicPRM is like a super-smart helper. Whenever you're stuck, it tells you what to do next and corrects your mistakes. This helper decides whether a correction is needed based on the importance of each step, just like in a game where some clues are more important than others. This way, EpicPRM helps you solve the puzzle faster without wasting too much time.

Glossary

EpicPRM

EpicPRM is a framework for constructing mathematical reasoning datasets, combining adaptive binary search and contribution quantification methods to improve annotation efficiency and data quality.

Used to construct the Epic50k dataset, enhancing PRM performance.

Adaptive Binary Search Algorithm

An algorithm that dynamically adjusts the starting position and sampling count to improve annotation efficiency and precision.

Used to identify the first erroneous step in reasoning chains.

Contribution Quantification Method

A method that improves annotation precision by quantifying the contribution of intermediate steps.

Used to annotate intermediate steps in reasoning chains.

PRM

A process-supervised reward model used to guide the reasoning process of large language models, improving reasoning abilities through step-by-step supervision.

Trained on Epic50k to enhance performance.

Epic50k

A high-quality mathematical reasoning dataset containing 50k annotated intermediate steps.

Used to train PRMs to improve reasoning performance.

Open Questions Unanswered questions from this research

  • 1 How can the EpicPRM framework be applied to a wider range of domains?
  • 2 How can the computational resource requirements of EpicPRM be optimized when handling more complex problems?

Applications

Immediate Applications

Mathematical Reasoning Training

The EpicPRM framework can be used to construct high-quality mathematical reasoning datasets suitable for training large language models.

Long-term Vision

Cross-Domain Applications

In the future, the EpicPRM framework may be applied to reasoning tasks in more domains, providing broader application scenarios.

Abstract

Enhancing the mathematical reasoning capabilities of Large Language Models (LLMs) is of great scientific and practical significance. Researchers typically employ process-supervised reward models (PRMs) to guide the reasoning process, effectively improving the models' reasoning abilities. However, existing methods for constructing process supervision training data, such as manual annotation and per-step Monte Carlo estimation, are often costly or suffer from poor quality. To address these challenges, this paper introduces a framework called EpicPRM, which annotates each intermediate reasoning step based on its quantified contribution and uses an adaptive binary search algorithm to enhance both annotation precision and efficiency. Using this approach, we efficiently construct a high-quality process supervision training dataset named Epic50k, consisting of 50k annotated intermediate steps. Compared to other publicly available datasets, the PRM trained on Epic50k demonstrates significantly superior performance. Getting Epic50k at https://github.com/xiaolizh1/EpicPRM.

cs.CL cs.AI